1 /* 2 * linux/fs/jbd2/recovery.c 3 * 4 * Written by Stephen C. Tweedie <sct@redhat.com>, 1999 5 * 6 * Copyright 1999-2000 Red Hat Software --- All Rights Reserved 7 * 8 * This file is part of the Linux kernel and is made available under 9 * the terms of the GNU General Public License, version 2, or at your 10 * option, any later version, incorporated herein by reference. 11 * 12 * Journal recovery routines for the generic filesystem journaling code; 13 * part of the ext2fs journaling system. 14 */ 15 16 #ifndef __KERNEL__ 17 #include "jfs_user.h" 18 #else 19 #include <linux/time.h> 20 #include <linux/fs.h> 21 #include <linux/jbd2.h> 22 #include <linux/errno.h> 23 #include <linux/slab.h> 24 #include <linux/crc32.h> 25 #endif 26 27 /* 28 * Maintain information about the progress of the recovery job, so that 29 * the different passes can carry information between them. 30 */ 31 struct recovery_info 32 { 33 tid_t start_transaction; 34 tid_t end_transaction; 35 36 int nr_replays; 37 int nr_revokes; 38 int nr_revoke_hits; 39 }; 40 41 enum passtype {PASS_SCAN, PASS_REVOKE, PASS_REPLAY}; 42 static int do_one_pass(journal_t *journal, 43 struct recovery_info *info, enum passtype pass); 44 static int scan_revoke_records(journal_t *, struct buffer_head *, 45 tid_t, struct recovery_info *); 46 47 #ifdef __KERNEL__ 48 49 /* Release readahead buffers after use */ 50 static void journal_brelse_array(struct buffer_head *b[], int n) 51 { 52 while (--n >= 0) 53 brelse (b[n]); 54 } 55 56 57 /* 58 * When reading from the journal, we are going through the block device 59 * layer directly and so there is no readahead being done for us. We 60 * need to implement any readahead ourselves if we want it to happen at 61 * all. Recovery is basically one long sequential read, so make sure we 62 * do the IO in reasonably large chunks. 63 * 64 * This is not so critical that we need to be enormously clever about 65 * the readahead size, though. 128K is a purely arbitrary, good-enough 66 * fixed value. 67 */ 68 69 #define MAXBUF 8 70 static int do_readahead(journal_t *journal, unsigned int start) 71 { 72 int err; 73 unsigned int max, nbufs, next; 74 unsigned long long blocknr; 75 struct buffer_head *bh; 76 77 struct buffer_head * bufs[MAXBUF]; 78 79 /* Do up to 128K of readahead */ 80 max = start + (128 * 1024 / journal->j_blocksize); 81 if (max > journal->j_maxlen) 82 max = journal->j_maxlen; 83 84 /* Do the readahead itself. We'll submit MAXBUF buffer_heads at 85 * a time to the block device IO layer. */ 86 87 nbufs = 0; 88 89 for (next = start; next < max; next++) { 90 err = jbd2_journal_bmap(journal, next, &blocknr); 91 92 if (err) { 93 printk (KERN_ERR "JBD: bad block at offset %u\n", 94 next); 95 goto failed; 96 } 97 98 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize); 99 if (!bh) { 100 err = -ENOMEM; 101 goto failed; 102 } 103 104 if (!buffer_uptodate(bh) && !buffer_locked(bh)) { 105 bufs[nbufs++] = bh; 106 if (nbufs == MAXBUF) { 107 ll_rw_block(READ, nbufs, bufs); 108 journal_brelse_array(bufs, nbufs); 109 nbufs = 0; 110 } 111 } else 112 brelse(bh); 113 } 114 115 if (nbufs) 116 ll_rw_block(READ, nbufs, bufs); 117 err = 0; 118 119 failed: 120 if (nbufs) 121 journal_brelse_array(bufs, nbufs); 122 return err; 123 } 124 125 #endif /* __KERNEL__ */ 126 127 128 /* 129 * Read a block from the journal 130 */ 131 132 static int jread(struct buffer_head **bhp, journal_t *journal, 133 unsigned int offset) 134 { 135 int err; 136 unsigned long long blocknr; 137 struct buffer_head *bh; 138 139 *bhp = NULL; 140 141 if (offset >= journal->j_maxlen) { 142 printk(KERN_ERR "JBD: corrupted journal superblock\n"); 143 return -EIO; 144 } 145 146 err = jbd2_journal_bmap(journal, offset, &blocknr); 147 148 if (err) { 149 printk (KERN_ERR "JBD: bad block at offset %u\n", 150 offset); 151 return err; 152 } 153 154 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize); 155 if (!bh) 156 return -ENOMEM; 157 158 if (!buffer_uptodate(bh)) { 159 /* If this is a brand new buffer, start readahead. 160 Otherwise, we assume we are already reading it. */ 161 if (!buffer_req(bh)) 162 do_readahead(journal, offset); 163 wait_on_buffer(bh); 164 } 165 166 if (!buffer_uptodate(bh)) { 167 printk (KERN_ERR "JBD: Failed to read block at offset %u\n", 168 offset); 169 brelse(bh); 170 return -EIO; 171 } 172 173 *bhp = bh; 174 return 0; 175 } 176 177 178 /* 179 * Count the number of in-use tags in a journal descriptor block. 180 */ 181 182 static int count_tags(journal_t *journal, struct buffer_head *bh) 183 { 184 char * tagp; 185 journal_block_tag_t * tag; 186 int nr = 0, size = journal->j_blocksize; 187 int tag_bytes = journal_tag_bytes(journal); 188 189 tagp = &bh->b_data[sizeof(journal_header_t)]; 190 191 while ((tagp - bh->b_data + tag_bytes) <= size) { 192 tag = (journal_block_tag_t *) tagp; 193 194 nr++; 195 tagp += tag_bytes; 196 if (!(tag->t_flags & cpu_to_be32(JBD2_FLAG_SAME_UUID))) 197 tagp += 16; 198 199 if (tag->t_flags & cpu_to_be32(JBD2_FLAG_LAST_TAG)) 200 break; 201 } 202 203 return nr; 204 } 205 206 207 /* Make sure we wrap around the log correctly! */ 208 #define wrap(journal, var) \ 209 do { \ 210 if (var >= (journal)->j_last) \ 211 var -= ((journal)->j_last - (journal)->j_first); \ 212 } while (0) 213 214 /** 215 * jbd2_journal_recover - recovers a on-disk journal 216 * @journal: the journal to recover 217 * 218 * The primary function for recovering the log contents when mounting a 219 * journaled device. 220 * 221 * Recovery is done in three passes. In the first pass, we look for the 222 * end of the log. In the second, we assemble the list of revoke 223 * blocks. In the third and final pass, we replay any un-revoked blocks 224 * in the log. 225 */ 226 int jbd2_journal_recover(journal_t *journal) 227 { 228 int err; 229 journal_superblock_t * sb; 230 231 struct recovery_info info; 232 233 memset(&info, 0, sizeof(info)); 234 sb = journal->j_superblock; 235 236 /* 237 * The journal superblock's s_start field (the current log head) 238 * is always zero if, and only if, the journal was cleanly 239 * unmounted. 240 */ 241 242 if (!sb->s_start) { 243 jbd_debug(1, "No recovery required, last transaction %d\n", 244 be32_to_cpu(sb->s_sequence)); 245 journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1; 246 return 0; 247 } 248 249 err = do_one_pass(journal, &info, PASS_SCAN); 250 if (!err) 251 err = do_one_pass(journal, &info, PASS_REVOKE); 252 if (!err) 253 err = do_one_pass(journal, &info, PASS_REPLAY); 254 255 jbd_debug(1, "JBD: recovery, exit status %d, " 256 "recovered transactions %u to %u\n", 257 err, info.start_transaction, info.end_transaction); 258 jbd_debug(1, "JBD: Replayed %d and revoked %d/%d blocks\n", 259 info.nr_replays, info.nr_revoke_hits, info.nr_revokes); 260 261 /* Restart the log at the next transaction ID, thus invalidating 262 * any existing commit records in the log. */ 263 journal->j_transaction_sequence = ++info.end_transaction; 264 265 jbd2_journal_clear_revoke(journal); 266 sync_blockdev(journal->j_fs_dev); 267 return err; 268 } 269 270 /** 271 * jbd2_journal_skip_recovery - Start journal and wipe exiting records 272 * @journal: journal to startup 273 * 274 * Locate any valid recovery information from the journal and set up the 275 * journal structures in memory to ignore it (presumably because the 276 * caller has evidence that it is out of date). 277 * This function does'nt appear to be exorted.. 278 * 279 * We perform one pass over the journal to allow us to tell the user how 280 * much recovery information is being erased, and to let us initialise 281 * the journal transaction sequence numbers to the next unused ID. 282 */ 283 int jbd2_journal_skip_recovery(journal_t *journal) 284 { 285 int err; 286 journal_superblock_t * sb; 287 288 struct recovery_info info; 289 290 memset (&info, 0, sizeof(info)); 291 sb = journal->j_superblock; 292 293 err = do_one_pass(journal, &info, PASS_SCAN); 294 295 if (err) { 296 printk(KERN_ERR "JBD: error %d scanning journal\n", err); 297 ++journal->j_transaction_sequence; 298 } else { 299 #ifdef CONFIG_JBD2_DEBUG 300 int dropped = info.end_transaction - be32_to_cpu(sb->s_sequence); 301 #endif 302 jbd_debug(1, 303 "JBD: ignoring %d transaction%s from the journal.\n", 304 dropped, (dropped == 1) ? "" : "s"); 305 journal->j_transaction_sequence = ++info.end_transaction; 306 } 307 308 journal->j_tail = 0; 309 return err; 310 } 311 312 static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag) 313 { 314 unsigned long long block = be32_to_cpu(tag->t_blocknr); 315 if (tag_bytes > JBD2_TAG_SIZE32) 316 block |= (u64)be32_to_cpu(tag->t_blocknr_high) << 32; 317 return block; 318 } 319 320 /* 321 * calc_chksums calculates the checksums for the blocks described in the 322 * descriptor block. 323 */ 324 static int calc_chksums(journal_t *journal, struct buffer_head *bh, 325 unsigned long *next_log_block, __u32 *crc32_sum) 326 { 327 int i, num_blks, err; 328 unsigned long io_block; 329 struct buffer_head *obh; 330 331 num_blks = count_tags(journal, bh); 332 /* Calculate checksum of the descriptor block. */ 333 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size); 334 335 for (i = 0; i < num_blks; i++) { 336 io_block = (*next_log_block)++; 337 wrap(journal, *next_log_block); 338 err = jread(&obh, journal, io_block); 339 if (err) { 340 printk(KERN_ERR "JBD: IO error %d recovering block " 341 "%lu in log\n", err, io_block); 342 return 1; 343 } else { 344 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data, 345 obh->b_size); 346 } 347 } 348 return 0; 349 } 350 351 static int do_one_pass(journal_t *journal, 352 struct recovery_info *info, enum passtype pass) 353 { 354 unsigned int first_commit_ID, next_commit_ID; 355 unsigned long next_log_block; 356 int err, success = 0; 357 journal_superblock_t * sb; 358 journal_header_t * tmp; 359 struct buffer_head * bh; 360 unsigned int sequence; 361 int blocktype; 362 int tag_bytes = journal_tag_bytes(journal); 363 __u32 crc32_sum = ~0; /* Transactional Checksums */ 364 365 /* Precompute the maximum metadata descriptors in a descriptor block */ 366 int MAX_BLOCKS_PER_DESC; 367 MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t)) 368 / tag_bytes); 369 370 /* 371 * First thing is to establish what we expect to find in the log 372 * (in terms of transaction IDs), and where (in terms of log 373 * block offsets): query the superblock. 374 */ 375 376 sb = journal->j_superblock; 377 next_commit_ID = be32_to_cpu(sb->s_sequence); 378 next_log_block = be32_to_cpu(sb->s_start); 379 380 first_commit_ID = next_commit_ID; 381 if (pass == PASS_SCAN) 382 info->start_transaction = first_commit_ID; 383 384 jbd_debug(1, "Starting recovery pass %d\n", pass); 385 386 /* 387 * Now we walk through the log, transaction by transaction, 388 * making sure that each transaction has a commit block in the 389 * expected place. Each complete transaction gets replayed back 390 * into the main filesystem. 391 */ 392 393 while (1) { 394 int flags; 395 char * tagp; 396 journal_block_tag_t * tag; 397 struct buffer_head * obh; 398 struct buffer_head * nbh; 399 400 cond_resched(); 401 402 /* If we already know where to stop the log traversal, 403 * check right now that we haven't gone past the end of 404 * the log. */ 405 406 if (pass != PASS_SCAN) 407 if (tid_geq(next_commit_ID, info->end_transaction)) 408 break; 409 410 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n", 411 next_commit_ID, next_log_block, journal->j_last); 412 413 /* Skip over each chunk of the transaction looking 414 * either the next descriptor block or the final commit 415 * record. */ 416 417 jbd_debug(3, "JBD: checking block %ld\n", next_log_block); 418 err = jread(&bh, journal, next_log_block); 419 if (err) 420 goto failed; 421 422 next_log_block++; 423 wrap(journal, next_log_block); 424 425 /* What kind of buffer is it? 426 * 427 * If it is a descriptor block, check that it has the 428 * expected sequence number. Otherwise, we're all done 429 * here. */ 430 431 tmp = (journal_header_t *)bh->b_data; 432 433 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) { 434 brelse(bh); 435 break; 436 } 437 438 blocktype = be32_to_cpu(tmp->h_blocktype); 439 sequence = be32_to_cpu(tmp->h_sequence); 440 jbd_debug(3, "Found magic %d, sequence %d\n", 441 blocktype, sequence); 442 443 if (sequence != next_commit_ID) { 444 brelse(bh); 445 break; 446 } 447 448 /* OK, we have a valid descriptor block which matches 449 * all of the sequence number checks. What are we going 450 * to do with it? That depends on the pass... */ 451 452 switch(blocktype) { 453 case JBD2_DESCRIPTOR_BLOCK: 454 /* If it is a valid descriptor block, replay it 455 * in pass REPLAY; if journal_checksums enabled, then 456 * calculate checksums in PASS_SCAN, otherwise, 457 * just skip over the blocks it describes. */ 458 if (pass != PASS_REPLAY) { 459 if (pass == PASS_SCAN && 460 JBD2_HAS_COMPAT_FEATURE(journal, 461 JBD2_FEATURE_COMPAT_CHECKSUM) && 462 !info->end_transaction) { 463 if (calc_chksums(journal, bh, 464 &next_log_block, 465 &crc32_sum)) { 466 put_bh(bh); 467 break; 468 } 469 put_bh(bh); 470 continue; 471 } 472 next_log_block += count_tags(journal, bh); 473 wrap(journal, next_log_block); 474 put_bh(bh); 475 continue; 476 } 477 478 /* A descriptor block: we can now write all of 479 * the data blocks. Yay, useful work is finally 480 * getting done here! */ 481 482 tagp = &bh->b_data[sizeof(journal_header_t)]; 483 while ((tagp - bh->b_data + tag_bytes) 484 <= journal->j_blocksize) { 485 unsigned long io_block; 486 487 tag = (journal_block_tag_t *) tagp; 488 flags = be32_to_cpu(tag->t_flags); 489 490 io_block = next_log_block++; 491 wrap(journal, next_log_block); 492 err = jread(&obh, journal, io_block); 493 if (err) { 494 /* Recover what we can, but 495 * report failure at the end. */ 496 success = err; 497 printk (KERN_ERR 498 "JBD: IO error %d recovering " 499 "block %ld in log\n", 500 err, io_block); 501 } else { 502 unsigned long long blocknr; 503 504 J_ASSERT(obh != NULL); 505 blocknr = read_tag_block(tag_bytes, 506 tag); 507 508 /* If the block has been 509 * revoked, then we're all done 510 * here. */ 511 if (jbd2_journal_test_revoke 512 (journal, blocknr, 513 next_commit_ID)) { 514 brelse(obh); 515 ++info->nr_revoke_hits; 516 goto skip_write; 517 } 518 519 /* Find a buffer for the new 520 * data being restored */ 521 nbh = __getblk(journal->j_fs_dev, 522 blocknr, 523 journal->j_blocksize); 524 if (nbh == NULL) { 525 printk(KERN_ERR 526 "JBD: Out of memory " 527 "during recovery.\n"); 528 err = -ENOMEM; 529 brelse(bh); 530 brelse(obh); 531 goto failed; 532 } 533 534 lock_buffer(nbh); 535 memcpy(nbh->b_data, obh->b_data, 536 journal->j_blocksize); 537 if (flags & JBD2_FLAG_ESCAPE) { 538 *((__be32 *)nbh->b_data) = 539 cpu_to_be32(JBD2_MAGIC_NUMBER); 540 } 541 542 BUFFER_TRACE(nbh, "marking dirty"); 543 set_buffer_uptodate(nbh); 544 mark_buffer_dirty(nbh); 545 BUFFER_TRACE(nbh, "marking uptodate"); 546 ++info->nr_replays; 547 /* ll_rw_block(WRITE, 1, &nbh); */ 548 unlock_buffer(nbh); 549 brelse(obh); 550 brelse(nbh); 551 } 552 553 skip_write: 554 tagp += tag_bytes; 555 if (!(flags & JBD2_FLAG_SAME_UUID)) 556 tagp += 16; 557 558 if (flags & JBD2_FLAG_LAST_TAG) 559 break; 560 } 561 562 brelse(bh); 563 continue; 564 565 case JBD2_COMMIT_BLOCK: 566 /* How to differentiate between interrupted commit 567 * and journal corruption ? 568 * 569 * {nth transaction} 570 * Checksum Verification Failed 571 * | 572 * ____________________ 573 * | | 574 * async_commit sync_commit 575 * | | 576 * | GO TO NEXT "Journal Corruption" 577 * | TRANSACTION 578 * | 579 * {(n+1)th transanction} 580 * | 581 * _______|______________ 582 * | | 583 * Commit block found Commit block not found 584 * | | 585 * "Journal Corruption" | 586 * _____________|_________ 587 * | | 588 * nth trans corrupt OR nth trans 589 * and (n+1)th interrupted interrupted 590 * before commit block 591 * could reach the disk. 592 * (Cannot find the difference in above 593 * mentioned conditions. Hence assume 594 * "Interrupted Commit".) 595 */ 596 597 /* Found an expected commit block: if checksums 598 * are present verify them in PASS_SCAN; else not 599 * much to do other than move on to the next sequence 600 * number. */ 601 if (pass == PASS_SCAN && 602 JBD2_HAS_COMPAT_FEATURE(journal, 603 JBD2_FEATURE_COMPAT_CHECKSUM)) { 604 int chksum_err, chksum_seen; 605 struct commit_header *cbh = 606 (struct commit_header *)bh->b_data; 607 unsigned found_chksum = 608 be32_to_cpu(cbh->h_chksum[0]); 609 610 chksum_err = chksum_seen = 0; 611 612 if (info->end_transaction) { 613 printk(KERN_ERR "JBD: Transaction %u " 614 "found to be corrupt.\n", 615 next_commit_ID - 1); 616 brelse(bh); 617 break; 618 } 619 620 if (crc32_sum == found_chksum && 621 cbh->h_chksum_type == JBD2_CRC32_CHKSUM && 622 cbh->h_chksum_size == 623 JBD2_CRC32_CHKSUM_SIZE) 624 chksum_seen = 1; 625 else if (!(cbh->h_chksum_type == 0 && 626 cbh->h_chksum_size == 0 && 627 found_chksum == 0 && 628 !chksum_seen)) 629 /* 630 * If fs is mounted using an old kernel and then 631 * kernel with journal_chksum is used then we 632 * get a situation where the journal flag has 633 * checksum flag set but checksums are not 634 * present i.e chksum = 0, in the individual 635 * commit blocks. 636 * Hence to avoid checksum failures, in this 637 * situation, this extra check is added. 638 */ 639 chksum_err = 1; 640 641 if (chksum_err) { 642 info->end_transaction = next_commit_ID; 643 644 if (!JBD2_HAS_INCOMPAT_FEATURE(journal, 645 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)){ 646 printk(KERN_ERR 647 "JBD: Transaction %u " 648 "found to be corrupt.\n", 649 next_commit_ID); 650 brelse(bh); 651 break; 652 } 653 } 654 crc32_sum = ~0; 655 } 656 brelse(bh); 657 next_commit_ID++; 658 continue; 659 660 case JBD2_REVOKE_BLOCK: 661 /* If we aren't in the REVOKE pass, then we can 662 * just skip over this block. */ 663 if (pass != PASS_REVOKE) { 664 brelse(bh); 665 continue; 666 } 667 668 err = scan_revoke_records(journal, bh, 669 next_commit_ID, info); 670 brelse(bh); 671 if (err) 672 goto failed; 673 continue; 674 675 default: 676 jbd_debug(3, "Unrecognised magic %d, end of scan.\n", 677 blocktype); 678 brelse(bh); 679 goto done; 680 } 681 } 682 683 done: 684 /* 685 * We broke out of the log scan loop: either we came to the 686 * known end of the log or we found an unexpected block in the 687 * log. If the latter happened, then we know that the "current" 688 * transaction marks the end of the valid log. 689 */ 690 691 if (pass == PASS_SCAN) { 692 if (!info->end_transaction) 693 info->end_transaction = next_commit_ID; 694 } else { 695 /* It's really bad news if different passes end up at 696 * different places (but possible due to IO errors). */ 697 if (info->end_transaction != next_commit_ID) { 698 printk (KERN_ERR "JBD: recovery pass %d ended at " 699 "transaction %u, expected %u\n", 700 pass, next_commit_ID, info->end_transaction); 701 if (!success) 702 success = -EIO; 703 } 704 } 705 706 return success; 707 708 failed: 709 return err; 710 } 711 712 713 /* Scan a revoke record, marking all blocks mentioned as revoked. */ 714 715 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh, 716 tid_t sequence, struct recovery_info *info) 717 { 718 jbd2_journal_revoke_header_t *header; 719 int offset, max; 720 int record_len = 4; 721 722 header = (jbd2_journal_revoke_header_t *) bh->b_data; 723 offset = sizeof(jbd2_journal_revoke_header_t); 724 max = be32_to_cpu(header->r_count); 725 726 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT)) 727 record_len = 8; 728 729 while (offset + record_len <= max) { 730 unsigned long long blocknr; 731 int err; 732 733 if (record_len == 4) 734 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset))); 735 else 736 blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset))); 737 offset += record_len; 738 err = jbd2_journal_set_revoke(journal, blocknr, sequence); 739 if (err) 740 return err; 741 ++info->nr_revokes; 742 } 743 return 0; 744 } 745